US7190723B2 - Digital stream transcoder with a hybrid-rate controller - Google Patents

Abstract

A rate controller in a transcoder, which receives a stream of compressed frames carried in a bit stream, selectively determines whether to quantize and/or threshold slices of a frame carried in the stream of frames. The rate controller determines the input size of the frame and based at least in part upon at least a desired size, requantizes and/or thresholds the frame such that the output size of the frame is approximately the desired size.

Description

CROSS-REFERENCE TO RELATED APPLICATION

This application is a continuation-in-part of U.S. application Ser. No. 10/635,406, filed on Aug. 6, 2003, which is a continuation-in-part of U.S. application Ser. No. 10/397,658, filed Mar. 26, 2003, which claimed priority to U.S. provisional application having Ser. No. 60/368,068, filed Mar. 27, 2002, all three of which are entirely incorporated herein by reference.

TECHNICAL FIELD

The present invention is generally related to broadband communication systems, and, more particularly, is related to transcoding compressed streams of information in broadband communication systems.

BACKGROUND OF THE INVENTION

Modern subscriber television systems (STS) transmit digital content, which is packetized, from a headend to a subscriber. The digital content is typically provided in a format such as MPEG or in other packet formats known to those skilled in the art. An operator of an STS typically prefers to provide programs in digital format because digital programs provide superior fidelity and because digital programs are compressed so that they generally use less bandwidth than analog programs. Digital programs are compressed using, in part, a quantization parameter.

Frequently, the operator of an STS may want to convert a compressed digital signal of a given bit rate into a compressed digital signal of a lower bit rate by using a conventional transcoder to change the quantization parameter. A conventional transcoder used for such a purpose consists of a cascaded decoder and encoder. This combination is rather complex and expensive. In the particular case of video signals, some other aspects have to be taken into account. A coded video signal consists of a succession of encoded video-frames, where each video-frame is subdivided into a two-dimensional array of macroblocks, each macroblock being composed of blocks. A video-frame may be in the spatial domain, which is the pixel domain, and is transmitted in the frequency or transform domain, which results from a Discrete Cosine Transform (DCT) of the video-frame in the spatial domain. In addition, a video-frame may be separated into two fields: the top field formed by the odd lines of the video-frame and the bottom field formed by the even lines of the video-frame. A macroblock may be conveyed in two different formats: an interlaced format and a de-interlaced format. In the interlaced video-frame format, a macroblock is composed of lines from the two alternating fields and each DCT-block of the macroblock is formed by data from the two fields. In the de-interlaced format, a macroblock is composed of lines from the two fields, and each DCT-block of the macroblock is formed by data from only one of the two fields. Each DCT-block of a video-frame is scanned and encoded.

Before a conventional pixel-domain transcoder can requantize a bit stream, the decoder portion of the transcoder converts the bit stream into pixel domain values. The encoder portion of the transcoder then requantizes and converts the pixel domain values back into DCT-domain values.

In addition to conventional pixel-domain transcoders, there exist conventional DCT-block domain transcoders, which operate in the DCT-block domain. Such a transcoder receives a bit stream and converts the bit stream into sets of run-level pairs, where a set of run-level pairs is a compressed representation of a DCT-block, and then converts the sets of run-level pairs into DCT-blocks. The transcoder manipulates information in the DCT-block domain and then reconverts the DCT-blocks back into sets of run-level pairs, which are then converted back into a compressed bit stream. Further details regarding DCT-block domain transcoders can be found in “A Frequency-Domain Transcoder For Dynamic Bit-Rate Reduction of MPEG-2 Bit Streams,” Assuncao et. al., IEEE Transactions on Circuits and Systems for Video Technology, Vol. 8,Issue 8, December 1998, pages 953–967, which is hereby incorporated by reference in its entirety; and “Manipulation and Compositing of MC-DCT Compressed Video,” Chang et al., IEEE Journal on Selected Areas In Communications, Vol. 13, No. 1, 1995,pages 1–11, which is hereby incorporated by reference in its entirety.

There exists a need for a transcoder that reduces the bit size of a stream such that the reduced bit size is approximately equal to a desired size, and a need for a method of reducing content so as to reduce the adverse effects of content reduction.

BRIEF DESCRIPTION OF THE DRAWINGS

The preferred embodiments of the invention can be better understood with reference to the following drawings. The components in the drawings are not necessarily to scale, emphasis instead being placed upon clearly illustrating the principles of the present invention. Moreover, in the drawings, like reference numerals designate corresponding parts throughout the several views.

FIG. 1 is a block diagram of a broadband communications system, such as a subscriber television system, in which the preferred embodiment of the present invention may be employed.

FIGS. 2A and 2B are illustrative pictures from a sequence of pictures.

FIG. 3 is a partial picture of the picture illustrated inFIG. 2B.

FIG. 4 is a residual picture.

FIG. 5 is a block diagram of a motion compensated block.

FIG. 6 is a block diagram of an encoder.

FIGS. 7A and 7B are diagrams of zig-zag scan order.

FIG. 8A is a diagram of a quantized matrix.

FIG. 8B is a diagram of a set of run-level pairs for the quantized matrix illustrated inFIG. 8A.

FIG. 8C is a diagram of a set of run-level pairs for the quantized matrix illustrated inFIG. 8A.

FIG. 9 is a block diagram of an embodiment of a transcoder

FIG. 10 is a graph of bit saving versus requantization parameter.

FIG. 11 is a diagram of a threshold function.

FIG. 12 is a block diagram of a rate controller.

FIG. 13 is a flow chart of steps taken implementing requantization/thresholding.

FIG. 14 is a flow chart of steps taken to determine whether to requantize.

FIG. 15 is a flow chart of steps taken to threshold.

FIG. 16 is a block diagram of states of a threshold state machine.

FIG. 17 is a block diagram of another embodiment of a transcoder.

FIG. 18 is a flow chart of steps taken in requantizing and thresholding a digital stream.

FIG. 19 is a flow chart of steps taken in motion compensation.

FIG. 20 is a flow chart of steps taken in accumulating drift.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

Preferred embodiments of the present invention will be described more fully hereinafter with reference to the accompanying drawings in which like numerals represent like elements throughout the several figures, and in which several exemplary embodiments of the invention are shown. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. The examples set forth herein are non-limiting examples and are merely examples among other possible examples.

Any process descriptions or blocks in flow charts should be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps in the process, and alternate implementations are included within the scope of the preferred embodiment of the present invention in which functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order, depending on the functionality involved, as would be understood by those reasonably skilled in the art of the present invention

One way of understanding the preferred embodiments of the invention includes viewing them within the context of a subscriber television system (STS). Thus, the preferred embodiments of the invention include, among other things, systems and methods for decreasing the size of transport streams carried by an STS.

Because the preferred embodiments of the invention can be understood in the context of a subscriber television system environment, an initial description of a subscriber television system (STS) is provided, which is then followed by a description of select components that are included within a headend of the subscriber television system. Also, a transcoder, which implements preferred embodiments of the invention and which is included in the headend at the headend, is described.

The preferred embodiments of the invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those having ordinary skill in the art. Furthermore, all “examples” given herein are intended to be non-limiting, and are provided as an exemplary list among many other examples contemplated but not shown.

Furthermore, it should be noted that the logic of the preferred embodiment(s) of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the preferred embodiment(s), the logic is implemented in software or firmware that is stored in a memory and that is executed by a suitable instruction execution system. If implemented in hardware, as in an alternative embodiment, the logic can be implemented with any or a combination of the following technologies, which are all well known in the art: a discrete logic circuit(s) having logic gates for implementing logic functions upon data signals, an application specific integrated circuit (ASIC) having appropriate combinational logic gates, a programmable gate array(s) (PGA), a field programmable gate array (FPGA), a digital signal processor (DSP) etc. In addition, the scope of the present invention includes embodying the functionality of the preferred embodiments of the present invention in logic embodied in hardware or software-configured mediums.

Subscriber Television System

FIG. 1 is a block diagram depicting a non-limiting example of a subscriber television system (STS)100. In this example, theSTS100 includes aheadend102, anetwork104, and multiple digital subscriber communication terminals (DSCTs)106, which are located atsubscriber premises105.

It will be appreciated that theSTS100 shown inFIG. 1 is merely illustrative and should not be construed as implying any limitations upon the scope of the preferred embodiments of the invention. For example, theSTS100 can feature a plurality of any one of the illustrated components, or may be configured with alternative embodiments for any one of the individual components or with yet other additional components not enumerated above. Subscriber television systems also included within the scope of the preferred embodiments of the invention include systems not utilizing physical structured cabling for transmission, such as, but not limited to, satellite systems.

ADSCT106, which is located at a subscriber'spremises105, provides among other things, a two-way interface between theheadend102 of theSTS100 and the subscriber. TheDSCT106 decodes and further processes the signals for display on a display device, such as a television set (TV)107 or a computer monitor, among other examples. Those skilled in the art will appreciate that in alternative embodiments the equipment for first decoding and further processing the signal can be located in a variety of equipment, including, but not limited to, a computer, a TV, a monitor, or an MPEG decoder, among others.

At least onecontent provider108 provides theSTS100 with digital content, which is formatted in a protocol such as, but not limited to, MPEG. Among other things, acontent provider108 can be a television station that provides “live” or “recorded” programming. A television station will include acamera110 and anencoder112. Theencoder112 receives content from thecamera110 and processes the content into an MPEG format, which is then provided to theheadend102 of theSTS100.

Theheadend102 receives programming signals from thecontent providers108, and, after processing the content from thecontent providers108 according to mechanisms described hereinbelow, theheadend102 transmits programming signals to theDSCTs106 at thesubscriber premises105. Typically, theheadend102 transmits a combination of both conventional analog signals (which will not be discussed) and digital signals.

In one implementation, the digital signals are transmitted in MPEG format and embodiments of the present invention will be discussed in terms thereof. Specifically, embodiments of the present invention are described in terms of MPEG video-frames and video-fields. However, it is to be understood that describing embodiments of the present invention employing MPEG video-frames and video-fields is merely for exemplary and clarity purposes and is not a limitation on the scope of the present invention. The scope of the present invention is intended to extend to at least to all streams of quantized information. For the purposes of this disclosure a frame of information includes video-frames, top video-fields, bottom video-fields, and other predetermined blocks of information.

As shown inFIG. 1, selected components of theexample headend102 include acommunications interface114, a digital network control system (DNCS)116, a conditional access (CA)server118, a video-on-demand (VOD)server120, atransport stream transmitter122, a quadrature phase shift keying (QPSK)modem124, arouter126, aVOD pump128, and atranscoder134, which are connected via anEthernet130. It will be understood by those having ordinary skill in the art that theexemplary headend102 can include additional components, such as additional servers, switches, multiplexers, transport stream transmitters, among others, or can omit some of the shown selected components.

Among other things, theDNCS116 manages, monitors, and controls network elements and the broadcast of services provided to users. TheDNCS116 includes, among other modules, asubscriber database132 that includes information about the subscribers for such purposes as billing information and survey data, among others. TheDNCS116 also communicates with theconditional access server118 to provide for secure transmittal of content from theheadend102 to theDSCTs106.

TheCA server118 selectively provides “entitlements” to theDSCTs106 for the services and programming of theSTS100. In other words, among other things, theCA server118 determines whichDSCTs106 of theSTS100 are entitled to access a given instance of service or program and provides the selectedDSCTs106 with, among other things, the necessary keys and authorizations to access the given instance of service. In addition, theCA server118 informs theDNCS116 of the entitlements of each of theDSCTs106 in theSTS100 so that each subscriber can be properly billed. Furthermore, theCA server118 includes a database (not shown) that includes, among other things, long term keys, the public keys of theDSCTs106 and a private key for theCA server118. The CA server employs long-term keys, public and private keys to securely communicate with theDSCTs106.

TheCA server118 also provides encryption information to thetransport stream transmitter122 and to the selectedDSCTs106. Thetransport stream transmitter122 employs the encryption information to encrypt the content of a program and transmits modulated programming, among other things, to theDSCTs110 via thenetwork104.

TheQPSK modem124 is responsible for transporting the out-of-band IP (Internet protocol) datagram traffic between theheadend102 and theDSCT106. Data transmitted or received by theQPSK modem124 may be routed by theheadend router126. Among other things, theheadend router126 may be used to deliver upstream data to the various servers, such as theVOD server120.

Thetranscoder134 receives aninput bit stream136 that carries a stream of MPEG transport packets and transmits anoutput bit stream138. The bit size of theoutput bit stream138 is smaller than theinput bit stream136. Thetranscoder134 is adapted to receive operator input and, among other things, apply a hybrid requantization-thresholding scheme on the frames of a program carried by theinput bit stream136. The hybrid requantization-thresholding scheme is performed in the DCT domain and is done such that the frames are reduced in bit size.

MPEG Compression

Before describing thetranscoder134 in detail, a brief description of MPEG video compression is provided. Further details of MPEG compression and MPEG in general can be found in MPEG-1 standards (ISO/IEC 11172), the MPEG-2 standards (ISO/IEC 13818) and the MPEG-4 standards (ISO/IEC 14496) are described in detail in the International Organization for Standardization document ISO/IEC JTC1/SC29/WG11 N (June 1996 for MPEG-1, July 1996 for MPEG-2, and October 1998 for MPEG-4), which are hereby incorporated by reference.

FIGS. 2A and 2B represent twopictures202A and202B, respectively, in a sequence of pictures.MPEG 2 segments a picture into 16×16 blocks of pixels calledmacroblocks204, which inFIGS. 2A and 2B are labeled1–25. In an actual high quality National Television System Committee (NTSC) frame, there are approximately1350 macroblocks. Eachmacroblock204 has a predefined location in a picture. For example, the macroblock labeled “1” is in the bottom right hand corner of each frame. As will be described herein, eachmacroblock204 if further subdivided into multiple 8×8 blocks of pixel information, which for the purposes of this disclosure are referred to as submacroblocks. A horizontal sequence of macroblocks is called a slice, and a slice can extend across the entire width of picture or a fraction of a width.

Conceptually, anMPEG 2 encoded picture consists of content information and non-content information. For purposes of this disclosure, content information is defined as the information that corresponds to pixel values in a macroblock, and non-content information corresponds to everything else necessary for processing and decoding the picture. Non-content information is generally carried in headers, examples of which include, but are not limited to, picture header, slice header, and macroblock headers. Headers typically carry information about how the picture, or portion thereof, was processed, so that the picture can be decoded and viewed. Non-content information include quantization parameters (Q1), which were used by an encoder to quantize portions of the picture and which are used to unquantize the picture. As will be explained in detail hereinbelow, content information generally corresponds to a submacroblock, and submacroblock can be represented in either pixel domain, DCT domain or run level domain. The different domains are described hereinbelow.

Picture2A illustrates aplane206A, acloud208, and background sky (not shown). Theplane206A is inmacroblocks1,2,6, and7; thecloud208 is inmacroblocks8,9,13, and14; and the background sky is in all of themacroblocks1–25.Picture202B illustrates the scene a short time later. Inpicture202B, theplane206B is now inmacroblocks13,14,15, and20, and asecond plane210 is entering thepicture202B inmacroblock5.

FIG. 3 illustrates a predictedpicture302. The predictedpicture302 includes theplane206A inmacroblocks13,14,15, and20, and thecloud208 inmacroblocks8,9,13, and14. The predictedpicture302 is based upon information contained inpicture202A. Specifically, theplane206A is translated frommacroblocks1,2,6, and7 ofpicture202A intomacroblocks13,14,15, and20 of predictedimage302, and thecloud208 is similarly translated fromFIG. 2A. Themacroblocks1,2,6, and7 ofpicture202A are shown as dashed lines inFIG. 3. Of course, the orientation, lighting, shading and other optical characteristics ofplane206A do not exactly match the image ofplane206B. Thus, the predictedimage302 is only an estimation of thepicture202B. To compensate for the differences between the predictedpicture302 and theactual picture202B a residual picture, illustrated inFIG. 4 is generated. Theresidual picture402 is the difference between the predictedpicture302 and theactual picture202B. For example, the difference betweenplane206B and206A is illustrated as aresidual plane404. Adding theresidual plane404 to theplane206A generates theplane206B.

Macroblock5 ofresidual picture402 is an example of an intracoded macroblock. Thesecond plane210 cannot be predicted from thereference picture202A and consequently, does not appear in the predictedframe402.

MPEG compresses content information using temporal compression and spatial compression. Temporal compression involves using information from a reference frame, such aspicture202A, to generate a predictedframe402 using motion vectors. Any macroblock having content from a reference picture has at least one motion vector associated with it; the motion vectors are carried in the macroblock header of that block. The motion vector identifies a macroblock in the reference frame from which the content information is taken.

Normally, a macroblock from a reference frame does not exactly coincide with the image in the current frame. For example,FIG. 5 illustrates a common situation where macroblock502 receives information from four macroblocks504(1)–504(4). Each one of thereference macroblocks504 is translated to themacroblock502 and offset such that only a portion of each of the fourreference macroblock504 is used inmacroblock502.

MPEG-2 employs three types of pictures, I-picture, B-picture, and P-picture. I-pictures are pictures that are intra-coded, i.e., compressed using only spatial compression from that video-frame, which means that they are decompressed without reference to any other video-frame. B-pictures and P-pictures are pictures that are inter-coded, i.e., compressed using information from a reference picture such as an I-picture or a P-picture, and are also spatially compressed. P-pictures are “predicted” pictures using information from a previous reference picture, and B-pictures are “bi-directionally predicted” pictures using information from a previous reference picture and from a subsequent reference picture. In practice, a B-picture or a P-picture is not strictly an inter-coded picture, but is instead a combination of inter-coded macroblocks and intra-coded macroblocks. Macroblocks that can be predicted from reference pictures are inter-coded and those cannot be predicted are intra-coded. Each macroblock has a macroblock header associated with it, and the macroblock header identifies the macroblock as being an inter-coded or intra-coded macroblock.

A typical sequence of video pictures in display order is I(1), B(2), B(3), P(4), B(5), B(6), P(7), B(8), B(9), P(10), . . . P(N), I(N+1). The P-picture P(4) uses information from the I-picture I(1); the B-pictures B(2) and B(3) use information from the I-picture I(1) and P-picture P(4); the P-picture P(7) uses information from the P-picture P(4); and the B-pictures B(5) and B(6) use information from the P-pictures P(4) and P(7). The pictures between I(1) and P(N), inclusive, are known as a group of pictures (GOP) and typically number between12–16, inclusive. Video pictures are not transmitted in display order. Instead, each inter-coded picture is transmitted after all of its reference pictures have been transmitted. Thus, the transmission order for a GOP is I(1), P(4), B(2), B(3), P(7), B(5), B(6), P(10), B(8), B(9), . . . P(N), B(N−2), B(N−1).

In a typical picture for display on a television, a high quality National Television System Committee (NTSC) frame is made up of approximately 1350 macroblocks. Common MPEG-2 standards include 4:2:0 and 4:2:2. In the 4:2:0 standard, a 16×16 macroblock is represented by a total of six sub-macroblocks (8×8): four 8×8 luminescent blocks; and two 8×8 color difference blocks, which are generated by down sampling each axis by a factor of 2. In the 4:2:2 standard, the chroma is not down sampled, and consequently there is twice as much chroma information. Thus, in the 4:2:2 standard, a 16×16 macroblock is represented by a total of eight sub-macroblocks. All of the sub-macroblocks of a macroblock are steered from a reference picture (I-picture or P-picture) to a temporally compressed picture (P-picture or B-picture) by a common motion vector.

Spatial compression in MPEG-2 is based upon transforming each sub-macroblock using a two dimensional discrete cosine transform (DCT) to convert from the pixel domain to the frequency domain, also known as the DCT domain. The steps in which an MPEG encoder, such asencoder112, spatially compresses frames are illustrated inFIG. 6. Theencoder112 includes atransformer602, aquantizer604, ascanner606, and abinary encoder608. Thetransformer602 transforms each sub-macroblock ofpixel information610 of a picture into aDCT domain sub-macroblock612 using a discrete cosine transform. Thepixel domain sub-macroblock610 is written as a matrix b, whose elements are given as b(n,m), where n and m range from 0 to 7, inclusive. TheDCT domain sub-macroblock612 is written as a matrix B, whose elements are given as B(k,j), where k and j range from 0 to 7, inclusive. Thetransformer602 uses the following equation to transform from pixel domain to DCT domain:

$\begin{array}{cc}\begin{array}{c}B\left(k,j\right)=\frac{c\left(k\right)}{2}\frac{c\left(j\right)}{2}\sum _{n=0}^7\sum _{m=0}^{7}b\left(n,m\right)\cos \left(\frac{\left(2n+1\right)·k\pi }{16}\right)\cos \left(\frac{\left(2m+1\right)·j\pi }{16}\right),\\ \mathrm{where}c\left(0\right)=1/\sqrt{2}\mathrm{and}c\left(n\right)=1\mathrm{for}n>0.\end{array}& \left(1\right)\end{array}$

The zero-frequency (DC) component, B(0,0), is in the top left hand corner ofDCT domain matrix612 and the coefficient for the highest frequencies, B(7,7), is in the bottom right hand corner of theDCT domain matrix612.

The DCT coefficients are not treated equally because the human eye is less responsive to high frequencies than low frequencies. Consequently, thequantizer604 applies a weight factor to each of the DCT coefficients while quantizing them. Quantization converts the DCT coefficients from rational numbers into integers and usually results in a sparse representation of the quantized DCT coefficients, i.e., one in which most or a large percentage of the amplitudes of the coefficients are equal to zero. In one implementation, thequantizer604 employs the following weight-quantization scheme:
B′(k,j)=int([2B(k,j)+1]·Q·w(k,j)/16,  (2a)
for inter-coded blocks and
B′(k,j)=int(2B(k,j)·Q·w(k,j)/16,  (2b)
for intra-coded block, where int( ) is the integer function, w(k,j) is the weight factor for element (k,j), and Q is the quantization parameter. An MPEG decoder would then employ the following inverse weight-quantization scheme:
B(k,j)=nint(B′(k,j)·16·Q/w(k,j)),  (3)
where nint( ) is the nearest integer function. Those skilled in the art recognize that other quantization schemes, which will not be discussed, but are intended to within the scope of the invention, can also be used.

Thescanner606 performs a zig-zag scan on the quantized DCT matrix (B′)614 and produces a run-level domain matrix (RL)616, which has the dimensions of (N+1)×2, where N is the number of non-zero coefficients in the quantized DCT matrix (B′)614. Finally, a binary encoder, or a variable length encoder (VLE),608 converts the run-level pairs of the run-level domain matrix (RL)616 into a bit stream using Huffman coding. It should be remembered that the preferred embodiments of the invention are being described in terms of MPEG standards, which use Huffman coding. However, the present invention is not intended to be limited to only MPEG standards and other coding techniques known to those skilled in the art can be uses in other preferred embodiments.

FIGS. 7A and 7B illustrate two possible scan orders. The scan order illustrated inFIG. 7A is typically implemented by thescanner606 for scanning the quantized DCT matrix (B′)614 when the DCT matrix represents a portion of a non-interlaced video-frame.FIG. 7B illustrates the scan pattern that is typically implemented when the DCT matrix represents a portion of interlaced video-fields.

FIG. 8A illustrates an exemplary quantized DCT-domain matrix (B′)614, andFIG. 8B illustrates the corresponding run-level domain matrix (RL)616 after thescanner606 has employed the scan pattern illustrated inFIG. 7A on the exemplary DCT-domain matrix614. In the run-level domain, “run” refers to the number of consecutively scanned coefficients having the value of zero that precede a non-zero coefficient, and “level” refers to the amplitude of the non-zero coefficients. The number of coefficients having the value of zero preceding the zero-frequency (D.C.) coefficient (B(0,0)=a) is zero, and thus the run-level pair for the D.C. coefficient is (0, a). The only zero coefficient interposing B(0,0) and B(1,0) is B(0,1), and thus the run-level pair for B(1,0) is given by (1, b). All of the coefficients following the B(4,1) coefficient (B(4,1)=h) are zero and are represented by an end-of-block marker, denoted by the run-level pair (0,0). Thus, after processing by thequantizer604 and thescanner606, the64 (rational number) coefficients in the DCT domain matrix (B)612 are now represented by nine pairs of runs and levels (18 integers). The conversion of 64 numbers into 18 integers (levels) reduces the number of bits necessary to represent the exemplary DCT-domain matrix614.

FIG. 8C illustrates an alternative embodiment of a set of run-level pairs616. In intra-coded macroblocks, MPEG-2 treats DC levels, the B(0,0) element ofmatrix614, differently from the higher frequency levels. The DC level of an intra block is encoded separately from the AC coefficients since the DC coefficient is differentially coded from block to block and because the human eye is more responsive to lower frequencies. Thus, there is no run value associated with the DC level because by definition that run would have to be zero. Whereas, in an inter-coded block, all of the levels in a block including the DC level are treated the same.

An MPEG decoder such as theDSCT106 performs inverse operations to convert a bit stream into frames. The MPEG decoder has a binary decoder (not shown) that, among other things, converts a bit stream into sets of run-level pairs, where a set of run-level pairs represents sub-macroblock of pixels. An inverse scanner (not shown) converts sets of run-level pairs into 8×8 matrices of DCT quantized coefficients. An inverse quantizer (not shown) multiplies the levels by the quotient of the quantization factor (Q) divided by the weight factor for each of the levels. Lastly, an inverse transformer (not shown) transforms the levels back into pixel domain values. Thus, MPEG encoding and decoding involve a lot of computational complexity due to, among other things, the matrix operations and DCT transformation and inverse transformations.

Transcoder

Illustrated inFIG. 9 are components of a first embodiment of thetranscoder134, andFIG. 17 illustrates components of a second embodiment. Referring toFIG. 9, thetranscoder134 includes a vector length decoder902 (VLD) aprocessor904 having amemory908, and a vector length encoder906 (VLE). Among other things, theVLD902 receives theinput stream136 and parses headers such as the picture headers, slice headers, macroblock headers, and others from the bit stream and provides the headers to thememory908. In addition, theVLD902 also parses non-video frames of information and provides the non-video frames to thememory908 and parses sets of run level pairs from the bit stream and provides the said run level pairs to theprocessor904.

Theprocessor904 processes frames so that, among other things, a processed frame is represented by fewer bits. The frames, video frames and non-video frames, are processed such that they are transmitted via theVLE906 in the same order in which they were received by theVLD902.

After processing a frame of information, theprocessor904 sends the processed frame to theVLE906. Among other things, theVLE906 converts the processed frame into binary information and encapsulates the binary information into multiple MPEG packets. The VLE converts run level pairs from pairs of integer values into binary sequences using well-known techniques, such as, but not limited to, Huffman coding.

Thememory908 has multiple buffers such as,reference frame buffers910A and910B, and shavedreference frame buffers912A and912B, in which reference frames and corresponding shaved reference frames are buffered, respectively. For the purposes of this disclosure, a shaved reference frame is one in which the bit size of the frame has been reduced. Thememory908 also includes buffers for non-video frames of information and for headers of the video frame.

Functionally, theprocessor904 can be thought of as being a cascaded encoder and decoder, which are separated by thedash line914. Aninverse quantizer module916, andinverse DCT module918, andadder920, and reference frame buffers910 make up the decoder portion, and anadder module922, aDCT module924, arate controller module926, aninverse quantizer module928, aninverse DCT module930, anadder module932, and the reference frame buffers912 make up the encoder portion.

The decoder portion of theprocessor904 converts a frame from the run level domain into pixel domain. Theinverse quantizer module916 receives content information as sets of run level pairs and inverse quantizes the levels in the sets based upon the initial quantization parameters (Q1), which are carried in one or more of the headers of the frame. Theinverse quantizer916 expands the unquantized levels from the run level domain into the DCT domain, i.e., theinverse quantizer916 converts a set of run level pairs into an 8×8 matrix representation by inverse zigzag scanning, or equivalently converting the set of run level pairs into an array of 64 levels arranged in scan order.

Theinverse DCT module918 receives content information in the DCT domain and converts the unquantized levels from frequency information back into pixel information by applying the inverse direct cosign transform to the DCT domain information.

Theadder module920 receives pixel information from theinverse DCT module918. If the current frame is an I-picture, then the pixel information is complete. If, however, the current frame is a P-picture or B picture then the pixel information is incomplete. Using the current frame's motion vectors, the information that is missing from the current picture is received from the reference frame buffers910. Theadder module920 adds the information from reference buffers910 to the pixel information from theinverse DCT module918. The output of theadder module920 is a complete frame. If the current frame is a reference frame (I-picture or P-picture), the current frame is sent to both the reference frame buffer910 for use with subsequent frames and to theadder module922 of the encoder portion of the processor. B-pictures are only sent to theadder module922.

When the current frame is an I-picture, theadder module922 provides the current frame to theDCT module924. However, when a current frame is a B picture or P-picture, theadder module922 generates a residual picture, which is then provided to theDCT module924. Using the current frame's motion vectors, theadder module922 generates a residual picture by subtracting predicted information stored in the shaved reference buffer912 from the current frame. The predicted information corresponds to the missing information that theadder module920 received from the reference frame buffers910.

TheDCT module924 converts content information from the pixel domain into the DCT domain where the levels of frequency information are unquantized. Therate controller926 includes aquantizer934, and athresholder936. Therate controller926 implements thequantizer934 andthresholder936 to reduce the size of the current frame such that the compressed bit size of the current frame is approximately equal to a desired bit size (N_D). The desired bit size is generally a parameter that an operator of the STS has provided, or which can be provided by theDNCS116 or by a frame-layer rate control algorithm which determines the number of bits in each picture frame based upon a target bit rate set by an operator.

Therate controller926 determines the current compressed bit size of the current frame and determines the number of bits to shave (N_S) therefrom using logic described hereinbelow. Therate controller926 quantizes, or thresholds, or quantizes and thresholds the current frame such that the compressed bit size is reduced by approximately (N_S).

If the current frame is a reference frame, therate controller926 provides the shaved frame to theinverse quantizer928. Therate controller926 also provides the shaved frame to thescanner938, which converts the content information from the DCT domain into run level domain. Thescanner938 then provides the shaved frame to theVLE906, which converts the content information from run level domain to compressed format.

Theinverse quantizer928 receives shaved reference frames from therate controller926 and converts the content information from quantized values into unquantized values of frequency information. The content information, which is now unquantized, is provided to theinverse DCT module932, which converts the content information back into pixel domain information.

Theadder934 receives content information, which is now pixel domain, and uses motion vectors of the current frame to get missing information from the shaved reference frame buffers912. The output ofadder934 is a complete shaved reference frame, which is then buffered in shaved reference frame buffers912 for use with subsequent predicted frames, i.e., P-pictures and B-pictures. Before discussing therate controller926 in detail a brief description of why certain requantization parameters (Q2) are used and a description of thresholding is provided.

FIG. 10 is a graph of χ versus the requantization parameter Q₂, where χ is defined as the quotient of the total size of the representative frame after requantization (N_T(Q₂)) divided by the total size of the representative frame before requantization (N_T(Q₁)). In the region labeledzone1, the magnitude of Q₂increases from Q₁, which is the original quantization parameter, up to approximately α, which is equal to 31 if a linear quantization scale is used, and112 if a non-linear quantization scale is used for the picture. The rate of change of χ with respect to Q₂(dχ/dQ₂) is discontinuous at Q₂=α, β, δ, and ε and is approximately constant between each of the discontinuities. The region between Q₂=Q₁to Q₂=α is defined aszone1 and throughout this region there is only an approximate 15% reduction in the size of the requantized frame. In the region defined aszone2, which extends from Q₂=β to Q₂=δ, the requantized frame is reduced by approximately 60%–70%, and in the region defined aszone3, which extends outward from Q₂=ε, the requantized frame is reduced at least by approximately 75%. The results shown inFIG. 10 are for a representative frame. The actual amount of reduction can vary depending upon variables such as the content of the frame, the type of picture, and other variables. Even so,FIG. 10 illustrates that it is normally preferable to use a requantization parameter from zone2 (or zone3) as opposed tozone1, because requantization inzone1 does not produce a significant saving in size.

As those skilled in the art will recognize, as the requantization parameter Q₂is increased, information is lost due to the requantization, which results in a lower quality of picture for the viewer. Thus, a balance between picture quality and size must be struck by the choice of requantization parameter Q₂. Preferably, the requantization parameter Q₂is not chosen fromzone1 because such a parameter only reduces the size of the requantized frame by at most approximately 15%. Instead, it is preferable that thresholding is used for such small decreases in the size of the frame. If requantization is performed, then in one preferred embodiment, the requantization reduces the size of the current frame to approximately the desired size, N_D, and then thresholding is performed to further reduce the size such that the total size of the frame is even closer to the desired size.

FIG. 11 illustrates anexemplary threshold function1102, which is a staired function having scan index thresholds1108, which are labeled I(0) through I(2), andlevel thresholds1110A, which are labeled L(0) through L(2). Therate controller926 zeros levels that are beneath thethreshold function1102. The level labeled1106A, whose scan position is between the scan index thresholds I(0) and I(1), is zeroed because the absolute value oflevel1106A is less than the level threshold L(0), which extends between the scan index thresholds I(0) and I(1). On the other hand, thelevel1104A is not zeroed because its absolute value exceeds the level threshold L(0). Similarly, thelevel1104B is not zeroed, and thelevels1106B and1106C are zeroed. In one preferred embodiment, therate controller926 thresholds the levels of a portion of a frame in parallel. In this embodiment, all of the sets of run level pairs that make up the portion are each thresholded by the same threshold function. Conceptually, as will be described in detail hereinbelow, therate controller926 moves thethreshold function1102 horizontally and vertically so that the correct number of levels are zeroed such that the size of the portion is reduced by approximately the appropriate amount.

Rate Controller

Referring toFIG. 12, in addition to thequantizer934,thresholder936, andscanner938, therate controller926 includes amemory1202 having aVLC table buffer1204, an N-bits buffer1206, aframe buffer1208, a workingbuffer1210, arun buffer1212, and alevel buffer1214. As those skilled in the art know, Huffman coding translates specific pairs of runs and levels to predetermined codes, which are of variable length. The most common run level pairs have the shortest codes. Some possible pairs of runs and levels are not assigned specific codes, and such run level pairs are represented by 24-bits: a 6-bit escape sequence; a 6-bit run sequence; and a 12-bit level sequence. TheVLC table buffer1204 includes a VLC table that maps the run level pairs having codes to their codes. The N-bits table buffer includes a table that maps the VLC codes to the size of the codes. Thus, therate controller926 can determine the compressed size of a portion of the current frame by the following equation:

$\begin{array}{cc}{S}_{\mathrm{SIZE}}=\sum _{J=0}^{\mathrm{Ncoef}-1}{\mathrm{VLC}}_J+24\times N_{\mathrm{ESCAPE}},& \left(4\right)\end{array}$
where Ncoef is the number of run level pairs in the given portion of the frame; VLC_Jis the number of bits of the variable length code for the J^thrun level pair in the portion and is zero if the J^thrun level pair is not in the VLC table; and N_escape is the number of run level pairs in the portion that do not have variable length codes assigned thereto. For each run level pair in the portion of the frame, therate controller926 first uses the VLC table to determine whether the pair has a specific code associated therewith, and if so, uses the N-bits buffer to determine the size (VLC_J) of the specific code.

The current frame is buffered in theframe buffer1208. Therate controller926 copies the frame into the workingbuffer1210 when thequantizer934 or thethresholder936 works on the frame or a portion thereof. If thequantizer934 processes the current frame, the result is copied into theframe buffer1208. As will be explained in detail hereinbelow, therate controller926 iteratively processes the current portion until the compressed size of the portion is approximately equal to a target size. For each iteration, thethresholder936 copies the portion of the frame from theframe buffer1208 into the workingbuffer1210.

When therate controller926 receives the current frame, thescanner938 scans the DCT domain content information and determines the pairs of runs and levels for the frame. The runs and levels are buffered in therun buffer1212 andlevel buffer1214, respectively. The runs and levels are then used with the VLC table and the N-bits table to determine various quantities such as, but not limited to, the total compressed size (N_T) of the frame, the total compressed content size of the frame (C_T), and the compressed content size of portions of the frame (S_size) such as a slice. The total compressed content size (C_T) is the total size of all of the content information when compressed. The compressed content size of the portion of the frame (S_size) is defined as the total size of all of the content information in that portion when compressed.

In one preferred embodiment, therate controller926 parses the frame into portions, such as slices, and then processes the portions sequentially until the entire frame is processed. Preferably, therate controller926 is adapted to process the sub-macroblocks of each portion in parallel. Before processing a portion of the frame, the transcoder determines a desired bit size for theoutput transport stream138. The desired bit size of thetransport stream138 is determined from operator input received through a user interface (not shown), or alternatively, received from theDNCS116. From the user input, the transcoder determines a desired bit size (N_D) for the compressed frames. Therate controller926 determines the target number of bits to shave from the portion (N_shave). After processing the portion, therate controller926 recalculates the compressed content size of the portion and determines the number of bits saved (N_saved), which is the difference between the initial compressed content size and the final compressed content size of the portion. Therate controller926 then determines the reduction error (e) which is defined as the difference between the target number of bits to shave (N_shave) and the number of bits saved (N_saved), e=N_shave—N_saved. The reduction error is accumulated for each portion and the accumulated reduction error (E) is used in the determination of the number of bits to shave from subsequent portions. For the K^thportion of the frame, N_shave is given as:

$\begin{array}{cc}{N}_{\mathrm{SHAVE}}=S_{\mathrm{SIZE}}\times \frac{N_S}{C_T}+\frac{E}{N_{\mathrm{SLICE}}-K+1},& \left(5\right)\end{array}$
where Ssize is the initial compressed content size of the K^thportion; C_Tis the total compressed content size of the frame; N_Sis the total number of bits to shave from the frame; E is the accumulated reduction error for previously processed portions,portions1 through K−1; and Nslice is the number of portions in the frame. Therate controller926 also determines a reduction threshold (R_T), which is given as:

$\begin{array}{cc}{R}_T=\frac{N_{\mathrm{SHAVE}}\left(K\right)}{S_{\mathrm{SIZE}}\left(K\right)}& \left(6\right)\end{array}$
The reduction threshold is used in the determination of whether or not to requantize the levels.

FIG. 13 illustrates exemplary requantization-thresholding logic implemented by therate controller926. Instep1302, a frame is received by therate controller926. The frame is parsed and buffered inmemory1202. Therate controller926 implements the hybrid requantization-thresholding scheme on a portion-by-portion basis. For the sake of clarity, in the discussion hereinbelow, a portion will be considered a slice. However, it is to be understood that a slice is a non-limiting example of a portion of a frame, and as those skilled in the art will recognize, the slice is an arbitrary portion of a picture and other smaller or larger portions of a picture may be utilized and are within the scope and intent of the invention. For example, a media processor or digital signal processor may have an internal cache which limits the portion of the picture which can be processed using the techniques set forth below.

Therate controller926 initializes parameters that are used in processing the entire frame such as the accumulated reduction error (E), and picture-type (P_T), among others. The type of picture, I-picture, P-picture or B-picture, is determined from the picture header, which is stored inmemory908. During initialization, therate controller926 also determines the amount of bits that need to be shaved off the frame (N_S).

Instep1304, therate controller926 determines quantities such as the slice content size, S_SIZE, and the reduction threshold, R_T,, the amount of bits to shave from the slice (N_SHAVE), and initializes slice quantities such as N_—SAVED.

Instep1306, therate controller926 determines whether to requantize the slice. Generally, the decision whether or not to requantize is based at least in part upon a requantization threshold parameter (T) and the reduction threshold (R_T). The requantization threshold parameter (T) is provided to thetransponder134 by theDNCS116 or by an operator, or is computed by a frame-layer rate control algorithm. Typically, if R_Tis greater than T then the slice is requantized. Other factors such as picture type and/or the initial quantization parameters used in quantizing the slice, among others, may also be used in the determination on whether to requantize or not. If the decision is not to requantize, therate controller926 proceeds to step1312, otherwise, the rate controller proceeds to step1308.

Instep1308, therate controller926 requantizes the levels of the current slice, and instep1310, therate controller926 determines the number of bits saved by requantization. The scanner scans the sub-macroblocks of the slice and generates new sets of run-level pairs for the slice. The new sets of run-level pairs are buffered in therun buffer1212 andlevel buffer1214. Therate controller926 uses theVLC table buffer1204 to determine the new codes for the requantized run-level pairs and the N-bits buffer1206 to determine the number of bits for the codes. For the Kth slice of the current frame the number of bits saved is given by the following equation:

$\begin{array}{cc}\mathrm{N\_saved}=S_{\mathrm{SIZE}}-\left(\sum _{J=0}^{\mathrm{Ncoef}\left(K\right)-1}{\mathrm{VLC\_NEW}}_J+{\mathrm{N\_escape}}_{\mathrm{new}}\times 24\right)& \left(7\right)\end{array}$
where VLC_NEW_Jis the compressed bit size of the new j^thrun-level pair, which is zero if the new j^thrun-level pair is not one of the specific codes found in theVLC table buffer1204, and N_escape_newis the new number of run-level pairs in the slice that are not found in theVLC table buffer1204.

Next instep1312, therate controller926 determines whether to the threshold the slice. Typically, the thresholding decision is based at least upon the number of bits saved, N_saved, which was initialized to zero instep1304 and, if necessary, calculated instep1310. If the number of bits saved, N_saved, is greater than or equal to the amount of bits to shave, N_shave, from the slice, therate controller926 proceeds to step1318. On the other hand, if N_saved is less than N_shave, therate controller926 proceeds to step1314 and thresholds the slice. Further details of the thresholding are provided hereinbelow.

Next, instep1316, therate controller926 determines the amount of bits saved, N_saved. The amount of bits saved is the difference between the number of bits used to represent the slice in compressed format, e.g., using Huffman code, and the initial size of the slice in compressed format. Typically the amount of bits saved will not exactly match the desired number of bits to shave from a slice, and the difference from the two values is added to the accumulated reduction error (E).

Instep1318, therate controller926 determines whether all of the slices of the frame have been processed, and if so, returns to step1302. Otherwise, it returns to step1304 and processes the next slice in the current frame. The processing described hereinabove was described in terms of processing a slice of the frame.

Table 1 lists adjustable parameters, which are provided by theDNCS116, or the operator, that are used by therate controller926 in determining whether to requantize. The adjustable parameters include the requantization threshold parameter (T), which in the preferred embodiment is an array, a quantization threshold array QT, which is a function of picture type (P_T), and LMIN, which is parameter associated with the average of the absolute value of the levels in the slice.

TABLE 1
Parameter	Example Value
T(0)	0.30
T(1)	0.40
T(2)	0.50
T(3)	0.60
T(4)	0.70
QT(0,P_T)	n/a
QT(1, P_T)	7 for P_T = I or P Picture, 9 for P_T = B picture
QT(2, P_T)	9 for P_T = I or P Picture, 11 for P_T = B picture
QT(3, P_T)	12 for P_T = I or P Picture, 14 for P_T =B picture
L
min	1

FIG. 14 further illustratesexemplary steps1400 for determining whether to requantize the current frame implemented by therate controller926 instep1306. Instep1402, a requantization flag is set to the default position of “false”, and a counter, “J,” is initialized to zero. Next instep1404, therate controller926 determines whether the reduction threshold, R_T, is less than the requantization threshold parameter T(J) for J=0. If the condition R_T<T(0) is true, therate controller926 drops to step1418 and is finished, which in this case means that requantization is not performed because the reduction threshold is so small that the current frame will be reduced to approximately the desired size by thresholding only. On the other hand, if the condition R_T<T(0) is false, therate controller926 proceeds to step1406.

Instep1406, therate controller926 increments the counter J, and instep1408, therate controller926 determines whether all of the following conditions are true: (i) R_T<T(J); (ii) Q1MAX<Q₂(J,P_T); and (iii) LAVG>LMIN, where Q1MAX is the maximum quantization parameter that was used to requantize the DCT blocks corresponding to the sets of run level pairs that make up the slice, and LAVG is the average of the absolute value of the levels that make up the slice. When the average absolute level of the slice LAVG is equal to 1, this means that at least half the levels of the slice have an absolute level of 1. Therefore, requantization by a factor of 2Q₁will necessarily zero half or more of the levels of the slice. Thus, in this situation, it is preferable to use thresholding instead of requantization to reduce the size of the slice. Only if all three conditions are true does therate controller926 proceed to step1416. On the other hand, if at least one of the three conditions is false, therate controller926 proceeds to step1410 and increments the counter “J”. Instep1412, therate controller926 determines whether the counter J is less than 4. Therate controller926 loops oversteps1408,1410 and1412 until either all three conditions ofstep1408 are true or until J=4.

Instep1412, which is reached when J=4, therate controller926 determines whether the reduction threshold R_Tis greater than the requantization threshold parameter T(4). If so, therate controller926 proceeds to step1416 and sets the requantization flag to “true”. If the condition R_T>T(4) is not met, therate controller926 drops to thelast step1418 and is finished with the requantization flag still set to the default “false”. However, if therate controller926 reachedstep1416 from eitherstep1408 or1414, the requantization flag is set to “true,” and then therate controller926 drops to thelast step1418 and is finished.

Referring back tostep1408, the three conditions ofstep1408 are exemplary conditions for determining whether or not to requantize. The three conditions are used so that the various factors such as the maximum initialization quantization parameter and picture type are included in the decision along with the reduction threshold and the average of the absolute value of the levels of the slice. Those skilled in the art will recognize that the conditions listed hereinabove are non-limiting lists and that other conditions or more conditions or fewer conditions beyond those listed hereinabove for selectively determining whether to requantize can also be used.

In one preferred embodiment, the requantization parameter Q₂for a set of run-level pairs is typically chosen to be 2Q₁or 4Q₁, where Q₁is the initial quantization parameter for the set of run-level pairs. Choosing the requantization parameter Q₂to be either 2Q₁or 4Q₁is done for computational efficiency, and the determination of whether to use 2Q₁or 4Q₁is based at least in part on the desired size of the requantized frame. However, it should be noted that the choice of 2Q₁or 4Q₁is a matter of implementation, and in alternative embodiments, the requantization parameter Q₂can be any quantization parameter. Typically, the default position is for Q₂to equal2Q₁, but if the condition R_T>T(4), or some other predetermined value, is true, then the value of Q₂is chosen such that Q₂=4Q₁. By choosing the requantization parameter Q₂to be either 2Q₁or 4Q₁, the requantization parameter Q₂is chosen fromzones2 or3 ofFIG. 10, respectively. Furthermore, it should be remembered that each set of run-level pairs of the current slice may not have been quantized with the same initial quantization parameter, and in that case, each set of run-level pairs is requantized using a requantization parameter that is a multiple of its initial quantization parameter, preferably Q₂=2Q₁or 4Q₁. Alternatively, the entire slice can be requantized using a common requantization parameter such as Q₂=2Q1 max.

Refer toFIG. 15,steps1500 illustrate an exemplary method to threshold the levels of a slice. The method starts atstep1502. Instep1504, therate controller926 determines the approximate number of levels (N_thresh) that need to be zeroed so that the size of the slice will be approximately the desired size after thresholding. The following equation is used to determine N_thresh for the current slice of the current frame:

$\begin{array}{cc}\mathrm{N\_thresh}=\frac{\left(\mathrm{Ncoef}-R_Q\right)\times R_T\times A\left(\mathrm{Run\_avg}\right)}{S_{\mathrm{SIZE}}},& \left(8\right)\end{array}$
where, Ncoef is the number of levels in the Kth slice, R_Qis the number of levels that were zeroed by requantization, Run_avg is the average run value of the current slice, and A( ) is a weighting function having Run_avg as its argument. It should be noted that R_Qis initialized to zero instep1304, and if requantization is performed, R_Qis tabulated instep1308. The weighting function A( ) strengthens the relationship from bits to levels as a function of the run average in the slice. Typically, as the average of the runs increases, the applied weight changes. For example, for an average run of zero, the run level pairs are coded efficiently using VLC, and consequently, A(0) is empirically determined to be approximately in the range of 1.2. Whereas, when the average of the runs is four, the run level pairs are not efficiently coded using VLC, and in that case, A(4) is empirically determined to be approximately in the range of 0.8.

In one preferred embodiment, the weighting function A( ) is adjusted instep1316 based upon the actual bits saved by thresholding. This enables on-line learning/feedback of the weighting function A( ) as a function of the average of the runs.

Next, instep1506, thresholding parameters are initialized, and the levels of the slice are buffered.

Instep1508, therate controller926 performs thresholding on the levels of the slice based upon the current position of the threshold function. The rate controller determines the number of sub-macroblock (Nblocks) in the slice and applies the threshold function to each sub-macroblock in the slice. Therate controller926 determines which levels of each block are beneath the threshold function and zeros those levels.

Instep1510, therate controller926 adjusts the threshold function by moving it vertically or horizontally so that the number of zeroed levels are closer to the value of N_thresh, or therate controller926 determines not to adjust the threshold function.

Instep1512, therate controller926 determines whether it is done with thresholding. If therate controller926 is finished, the method ends instep1514. Otherwise, the method loops back tostep1508. Eachtime step1508 is entered, the levels of the slice are reset according to the buffered levels ofstep1506.

Typically, the number of levels that are set to zero by thresholding will not exactly be equal to the desired value of N_thresh or be within a predetermined range of the desired value of N_thresh. Thus, in one preferred embodiment, therate controller926 partitions the slice into a first group and a second group of sub-macroblocks. Therate controller926 then adjusts the threshold function for each group independently. If the total number of zeroed levels in the first and second group is still not within a predetermined range of N_thresh, therate controller926 transfers a predetermined number of sub-macroblocks from the second group into the first group. Therate controller926 continues to transfer sub-macroblocks from the second group into the first group, determine the number of threshold levels, and if the number of threshold levels is not within the predetermined range of N_thresh, transfer more sub-macroblocks from the second group to the first group until the total number of zeroed levels is within the predetermined range.

In one preferred embodiment, therate controller926 implements a state machine, the states of which are illustrated inFIG. 16, for adjusting the index and threshold levels of the threshold function. The state machine can be seen as passing through a level threshold search followed by a scan index threshold search, with states along the way. Those skilled in the art will recognize that the threshold function illustrated inFIG. 11 was an exemplary threshold function having three levels and that threshold functions having a different number of levels are intended to be within the scope of the present invention. For example, presently described hereinbelow, therate controller926 implements a four level threshold function. Parameters that are used by the state machine are initialized instep1506 and shown in Table 2.

	TABLE 2
	Parameter	Value
	L(0)	2
	I(0)	index_thresh_min
	SplitIndex Val
	0
	φ	0.05
	UL	(1 + φ) × N_thresh
	LL	(1 − φ) × N_thresh
	STATE	FINDING_LEVEL_POS
	α(K)	1 + 5 // QAVG
	offset1	8
	offset2	4
	offset3	6

The parameters are defined as follows:

L(0): level threshold ofindex segment0 labeled1110A inFIG. 11;

I(0): scan index threshold ofindex segment0 labeled1108A inFIG. 11;

SplitIndexVal: the number of blocks in the first group when split indexing is performed;

φ:	adjustable parameter for defining a thresholding windows;
UL:	upper limit on number of number of levels thresholded to zero;
LL:	lower limit on number of number of levels thresholded to zero;

α(K): α(K)=1+5//QAVG, where // denotes integer division with truncation and QAVG is the average of the initial quantization parameters (Q₁) of the slice, and the parameter α(K) is used for setting level threshold for indices greater than 0; and

offset_(1,2,3): tunable parameters used for setting index thresholds for indices greater than 0;

index_thresh_min: 0 for B-frame, 1 for I or P frame.

The threshold level (L) for the index segment zero of the threshold function is initialized to 2, and the remaining threshold levels of the threshold function are given as follows:
L(n)=L(n−1)+

,  (9)
where n ranges from 1 to three. The levels are incremented α(K). Because α(K) is a function of QAVG, the average of the initial quantization parameters (Q₁) of the slice, the rise in the level threshold from one index segment to the next is sensitive to the quantizer scale.

The scan index theshold I(0) for of the threshold function is initialized and held at index_thresh_min (I(0)=index_thresh_min) during the level search (states FINDING_LEVEL_POS and FINDING_LEVEL_NEG), and is initialized to ISTART at the start of the index search, when the state FAST_INDEX_SEARCH is entered, where ISTART is is given as follows:
ISTART=γ×(1−R_T)×IAVG(K)  (10)
where IAVG is the average scan position of the levels in the Kth slice and γ is a tunable parameter and which is approximately 2.75

For the remaining scan index thresholds n=1 through 3, I(n) is given as follows:
I(n)=I(n−1)+offset_n  (11)
where offset_nis specified in Table 2.

All scan index thresholds I(n) for n=0 through 3 are checked to make certain that they are less than or equal to 63 because the scan positions only run to 63. If I(n) is greater than 63, it is simply set to 63.

Referring toFIG. 16, the state machine can be seen as passing through a level threshold search followed by a scan index threshold search, with states along the way. Theinitial state1602 is FINDING_LEVEL_POS. InFIG. 16, conditional expressions are shown inside of dashed ellipses, and actions taken by the state machine are underlined.

State Finding _Level_Pos:

The purpose of theinitial state1602 is to increment the level threshold L(0) until the count of the thresholded levels (cnt) exceeds the target count (N_thresh), where cnt is the number of levels zeroed. In this state, the threshold function is not moved horizontally as the state machine attempts to determine the minimum threshold levels that satisfy cnt>N_thresh. Instead, I(0) is held at index_thresh_min and the lowest level threshold L(0) is incremented by α. The level thresholds L(1), L(2), and L(3) are recomputed as L(n)=L(n−1)+α, for n=1, 2, and 3, until the condition cnt>N_thresh is met. Typically, the levels of a set of run-level pairs are populated most densely around small scan-positions, and consequently, during the index search, the cnt will be backed off of (made lesser) by sliding the threshold function to the right, e.g., making I(0)>index_thresh_min and recalculating I(1)–I(3).

To limit the number of iterations through this state, a higher increment than α may be used after a predetermined number (IT) of unsuccessful iterations, where IT is a tunable parameter, e.g., IT=5. For example, if the number of iterations is greater IT, the threshold level for the index segment zero (L(0)) can be given as:
L(0)=L(0)+2×iterations.  (12)
Alternatively, a binary search can be employed. In most cases, especially in B-pictures and P-pictures where the sets of run-level pairs contain residual information, the final level threshold is often the initial guess of L(0)=2.

After the levels of the threshold function have been raised, if needed, such that the condition cnt>N_thresh is met, the height of the threshold level L(0) is considered to be a minimum if the last increment was α. In this case, the level threshold is final and the state machine moves to theFAST_INDEX_SEARCH state1606.

However, if instead it took a large number of iterations through this state to find L(0) and the last increment was not by α, then the threshold level L(0) is not a minimum. In this case, the state machine proceeds toFINDING_LEVEL_NEG state1604.

Finding_Level_Neg:

TheFINDING_LEVEL_NEG state1604 is entered after theFINDING_LEVEL_POS state1602 zeroed more than N_thresh levels and the last increment was more than α. Typically, this situation occurs when there is a high number of iterations and the increment for the levels is given byequation 12.

In this situation, the threshold level L(0) is not a minimum and theFINDING_LEVEL_NEG state1604 decrements L(0) by α, while holding the index threshold at index_thresh_min, until the condition cnt<N_thresh is met or until the threshold level L(0) is back to its initial value. If the condition cnt<N_thresh is met, then the threshold levels have been decremented too far, and in that case the threshold levels are incremented by α.

Fast_Index_Search:

The purpose of theFAST_INDEX_SEARCH state1606 is to quickly find the neighborhood that the final scan index threshold is in by incrementing or decrementing the scan index threshold by a coarse increment, for example, β=4. The initial scan index thresholds I(n) (n=0 . . . 3) were set instep1506. If cnt is less than the lower limit of the index window, LL, and the value of cnt on the last iteration of the state machine (last_cnt) was less than or equal to LL, then the index threshold I(0) is decreased by β. On the other hand, if cnt is greater than the upper limit, UL, and the preceding cnt (last_cnt) was greater than or equal to UL, then the index threshold I(0) for is increased by β.

If cnt is greater than UL, but the preceding cnt (last_cnt) was less than UL, then the fast index search went too far left (towards lower frequencies). In this case, the index threshold I(0) is incremented by β−1 and the state is modified to theMOVING_LEFT state1610.

If cnt is less than LL, but the preceding cnt (last_cnt) was greater than LL, then the fast index search went too far right (towards higher frequencies). In this case, the index threshold I(0) is decremented by β−1 and the state is modified to theMOVING_RIGHT state1608.

Moving_Right:

When in theMOVING_RIGHT state1608, the cnt is checked against UL. If (cnt>UL), then the scan index threshold I(0) is incremented by 1. If cnt becomes less than LL, then theMOVING_RIGHT state1608 went one index too far. In this case, the scan index threshold I(0) is decremented by 1, and the state machine proceeds to theSPLIT_INDEX state1612, where the SplitIndexVal is set to 1 block of levels.

If neither of the above conditions are satisifed, i.e. (LL<cnt<UL ), then the state machine proceeds to the DONE state1614, where state machine returns the state “Done” and stops.

Moving_Left:

When in theMOVING_LEFT state1610, the cnt is checked against UL. If (cnt>UL), then the scab index threshold I(0) is incremented by 1. If cnt becomes less than LL, then theMOVING_LEFT state1610 went one index too far. In this case, the index threshold I(0) is decremented by 1, and the state machine proceeds to theSPLIT_INDEX state1612, where the SplitIndexVal is set to 1 block of levels.

If neither of the two conditions above are met, i.e. (LL<cnt<UL), then the state machine proceeds to the DONE state1614, where state machine returns the state “Done” and stops.

Split_Index:

The SPLIT_INDEX state1712 splits (or segments) the levels of the slice into two segments as defined by SplitIndexVal, so that not all levels of the slice are handeled equally. The thresholding operations up until the state machine enters the SPLIT_INDEX state have SplitIndexVal=0, so there is no split index thresholding up until this point.

One reason for the SPLIT_INDEX state1712 is that thresholding at a particular value of I(0)=t, where t is determined by theMOVING_LEFT state1610 or theMOVING_RIGHT state1608, results in cnt>UL but thresholding with I(0)=t+1 results in cnt<LL. In this case, it is impossible to find a scan position for the index threshold I(0) such that cnt is within the window (LL<cnt<UL ). Therefore, in the first segment of levels the index threshold I(0) is set to t, and in the second segment of the index threshold I(0) is set to t+1. If the total cnt for both segments is less than UL, then the state machine proceeds to the DONE state1614, where state machine returns the state “Done” and stops. On the other hand, if the total cnt for both segments is not less than UL, then the SplitIndexVal is incremented so that more levels are moved from the first segment to the second segment. When cnt reaches the condition (cnt<UL), the state machine proceeds to the DONE state1614, where state machine returns the state “Done” and stops.

Hereinbelow is an exemplary pseudocode for performing split indexing over two partitions. The first partition runs from 0 to Nblocks-SplitIndexVal−1 and the second partition runs from Nblocks—SplitindexVal toNblocks−1. The parameter SplitIndexVal controls where the dividing line between the partitions is placed. This effectively gives some fine tuning when the count of the thresholded coefficients is too large (greater than UL) at one scan index threshold but is too small (less than LL) at the neighboring index one away. Therefore, when SplitIndexVal is set to non-zero, thresholding is done with the threshold function starting at scan index I(0) for the first partition and starting at scan index I(0)+1 for the second partition. SplitIndexVal is initialized to zero at the beginning of the slice thresholding and is modified by the state machine to move the count of thresholded coefficients within the window defined between LL and UL.

Set Rest of Level and Index Thresholds:
L(n) = L(n−1) + α   (1<n<3)
I(n) = I(n−1) + offsetn.(1<n<3)
Reset Levels (Level(j) to orginal values
Loop w over 4 thresholds [L(0) I(0)] to [L(3) I(3)]
{

	Loop i from 0 to Nblocks - SplitIndexVal - 1
	{

	If abs( Level(i) ) > L(n)  AND Scan-Position(i) > I(n) − 1
	{
	Level(i) = 0
	}

}

}

Loop w over 4 thresholds [L(0), I(0)] to [L(3) I(3)]

{

	Loop i from Nblocks - SplitIndexVal to Nblocks - 1
	{

	If abs( Level(i) ) > L(n)  AND Scan-Position(i) > I(n)
	{
	Level(i) = 0
	}

}

}

Second Embodiment

Referring toFIG. 17, in a second preferred embodiment of the invention, thetranscoder134 includes theVLD902, theVLE906, and aprocessor1702. TheVLD902 andVLE906 were previously described and shall not be described again. Furthermore, in the second preferred embodiment, therate controller1704 determines whether or not to requantize, or threshold, or requantize and threshold using the requantization/thresholding logic illustrated inFIGS. 13–16.

Theprocessor1702 includes arate controller1704,adder1706 andmemory1708. Thememory1708 includesdrift buffers1710A and1710B and other buffers (not shown) for, among other things, motion vectors, headers, non-video frames, the N-bits table, and the VLC table. In addition to thethresholder936, andscanner938, therate controller1704 includes amotion compensation module1712 and arequantizer module1714. In this embodiment, instead of applying the motion compensation in the pixel domain, theprocessor1702 applies motion compensation to in the DCT-domain,.

As will be explained in detail hereinbelow, translating a block of pixel information from a reference sub-macroblock into a current sub-macroblock is equivalent to multiplying the sub-macroblock (in matrix format) by window functions. Because the window functions are unitary orthogonal matrices, the DCT transform of the product of the window function times the sub-macroblock of pixels is distributive, and consequently, the product is equal to the matrix product of the DCT representation of the window function times the DCT representation of the sub-macroblock. The set of all possible motion vectors is finite and thememory1708 includes DCT domain motion compensation matrices, (G) that are used by the motion compensator. The drift buffers1710 have accumulated drift for each sub-macroblock for two reference frames stored therein, where the drift of a sub-macroblock is the difference between the unquantized levels of a sub-macroblock before processing and the unquantized levels after processing, i.e., after reducing the bit size of the levels by requantization and/or thresholding. Preferably, the drift of a sub-macroblock is stored in array format, or equivalently, it can also be stored in matrix format and can be mapped back and forth between the two formats.

Therate controller1704 receives content information included in a current frame, and thescanner938 converts the content information from run-level domain into DCT domain. In other words, thescanner938 expands each set of run level pairs into 64 levels, some or most of which are zero. The levels can be arranged in either an 8×8 matrix or a 64-element array. As will be explained in detail hereinbelow, it is preferable to arrange the levels of a sub-macroblock in scan order in a 64-element array and to accumulate the drift of sub-macroblocks in 64 element arrays.

Themotion compensator1712 receives accumulated drift (D) from the drift buffer1710 and uses motion vectors to select appropriate motion compensation matrices (G). The accumulated drift is matrix multiplied by the appropriate motion compensation matrix (G) and the product (GD) is added to the submacroblock of the current frame.

When an I-picture is received by therate controller1704, no motion compensation is performed. However, therate controller1704 includes buffers for the unquantized levels, which are denoted by (L), and the unquantized levels are provided to theadder1706. Therate controller1704 also provides theadder1706 with unquantized reduced levels, which are denoted by (L′). For a sub-macroblock, the unquantized reduced levels are the unquantized levels after the size of the macroblock has been reduced/shaved byrequantizer1704 and/or thethresholder936. The drift of a sub-macroblock in an I-Picture is the difference between the unquantized levels (L) before processing and the unquantized reduced levels (L′).

Theadder1706 provides the drift to thememory1708, which buffers the drift for the sub-macroblock. Once the memory has the drift for all of the sub-macroblocks of the current frame, the drift for the frame is stored in the drift buffer1710.

For each subsequent frame, therate controller1704 extracts drift from the drift buffer1710 and applies motion compensation to it and adds the motion compensated drift (GD) to the unquantized levels of the current frame: (L)=(L)+(GD), where (L) is a matrix/array of unquantized levels for a sub-macroblock of the current frame, D is a matrix/array of the accumulated drift for a reference sub-macroblock; and (G) is the motion compensation matrix associated with the motion vector for the sub-macroblock of the current frame. The motion compensated drift (GD) is also provided to theadder1706. Therate controller1704 requantizes/thresholds levels of the current frame and provides theadder1706 with both unquantized levels (L) and reduced unquantized levels (L′) of the current frame. The accumulated drift for a sub-macroblock is then given by the following equation:
D=(GD)+(I−I′)  (13)
After D′ has been calculated for all of the sub-macroblocks of the current frame, the accumulated drift of the current frame is buffered in the drift buffer1710.

As previously described an inter-coded frame is generated at an MPEG decoder by adding pixel information from blocks in a reference frame to pixels of a residual frame. The MPEG decoder uses motion vectors, which are included in the headers of the inter-coded frame, to translate a block of pixel values from a reference frame to the inter-coded frame. Typically, a motion compensated block, one in which information is retrieved from one or more reference frames, is made up of portions of more than one reference block.FIG. 5 illustrates a common situation, which occurs when both components of a motion vector are not integer multiples of the block size, e.g., 8 pixels. The motion compensatedblock502 is made up of foursub-blocks508, which are labeled 1–4, and a residual block (not shown). Each sub-block508 is a portion of the reference blocks504. Sub-block508(1) is (A×B) in size, where “A” is the number of rows of pixels and “B” is the number of columns of pixels, and corresponds to the bottom right hand corner of reference block504(1); sub-block508(2) is (A×(8-B)) in size and corresponds to the bottom left hand corner of reference block504(2); sub-block508(3) is ((8-A)×B) in size and corresponds to the top right hand corner of reference block504(3); and sub-block508(4) is ((8-A)×(8-B)) in size and corresponds to the top left hand corner of reference block504(4). Themotion vectors506, r₁–r₄, translate the reference blocks504(1)–504(4) such that the sub-blocks508(1)–508(4) are appropriately positioned.

In matrix form, the motion compensatedblock502 is denoted by d^mcand is given by the following equation:

$\begin{array}{cc}{d}^{mc}=\sum _{i=1}^4{d}_i,& \left(14\right)\end{array}$
where d_iis an 8×8 matrix given by the following equation:
d_i=h_i^nrb_iw_i^nc,  (15)
where b_iis the i^threference block504, nr and nc are the number of rows and columns, respectively, of the sub-block508(i), and h_i^nrand w_i^ncare of the form of upper and lower diagonal matrices having identity sub-matrices. The h matrices for the four sub-blocks508 are as follow:

$h_1^{nr}=\left[\begin{array}{cc}0& I^{n\mathrm{<figure-callout\; label="r"\; filenames="US07190723-20070313-D00004.png,US07190723-20070313-D00008.png"\; state="\{\{state\}\}">r</figure-callout>}}\\ 0& 0\end{array}\right],h_2^{nr}=\left[\begin{array}{cc}0& I^{n\mathrm{<figure-callout\; label="r"\; filenames="US07190723-20070313-D00004.png,US07190723-20070313-D00008.png"\; state="\{\{state\}\}">r</figure-callout>}}\\ 0& 0\end{array}\right],h_3^{nr}=\left[\begin{array}{cc}0& 0\\ I^{8-nr}& 0\end{array}\right],\mathrm{and}{h}_4^{nr}=\left[\begin{array}{cc}0& 0\\ I^{8-nr}& 0\end{array}\right];$
and the w matrices are as follows:

$w_1^{nc}=\left[\begin{array}{cc}0& 0\\ I^{nc}& 0\end{array}\right],w_2^{nc}=\left[\begin{array}{cc}0& I^{8-n\mathrm{<figure-callout\; label="c"\; filenames="US07190723-20070313-D00004.png,US07190723-20070313-D00008.png"\; state="\{\{state\}\}">c</figure-callout>}}\\ 0& 0\end{array}\right],{\mathrm{<figure-callout\; label="w"\; filenames="US07190723-20070313-D00002.png,US07190723-20070313-D00004.png"\; state="\{\{state\}\}">w</figure-callout>}}_3^{nc}=\left[\begin{array}{cc}0& 0\\ I^{nc}& 0\end{array}\right],\mathrm{and}{w}_4^{nc}=\left[\begin{array}{cc}0& I^{8-n\mathrm{<figure-callout\; label="c"\; filenames="US07190723-20070313-D00004.png,US07190723-20070313-D00008.png"\; state="\{\{state\}\}">c</figure-callout>}}\\ 0& 0\end{array}\right].$
Applying the discrete cosine transform toequation 15 yields:

$\begin{array}{cc}\mathrm{DCT}\left(d\right)=\sum _{i=1}^4\mathrm{DCT}\left(d_i\right)=\sum _{i=1}^4\mathrm{DCT}\left(h_i^{nr}\right)\mathrm{DCT}\left(b_i\right)\mathrm{DCT}\left(w_i^{nc}\right)& \left(16a\right)\\ D=\sum _{i=1}^4{D}_i=\sum _{i=1}^4{H}_i^{nr}{B}_i{W}_i^{nc},& \left(16b\right)\end{array}$
because the h_i^nrand w_i^ncmatrices are unitary orthogonal, the DCT operation is distributive. All of the matrices inequations 14–16 are 8×8 in size, and consequently, by arranging the elements of the D, D_i, and B_imatrices in a predetermined order, such as the scan order shown inFIG. 7A, each component of equation 16b can be rewritten as
D′_i=G_i(H_i^nr,W_i^nc)B′_i,  (17)
where the primed matrices are 64×1 in size and G, which is a function of the H_iand W_imatrices, is a 64×64 matrix that is calculated from the H_iand W_imatrices, and where the subscript “i” refers to the i^threference block. As shown inFIG. 5, “i” normally runs from 1–4. However, for the sake of clarity the subscript “i” will be dropped, which means that the magnitude of the components of the motion vector, which extends from the reference frame to the current frame, are each an integral number of blocks.

Consider matrices a, b, c, d and e, which are all the same size (N×N), where
a=cbd=(cb)d=ed,  (18)
and the (n,m) component of matrix a is given by

$\begin{array}{cc}{a}_{n,m}=\sum _{\alpha =0}^{N-1}{e}_{n,\alpha }{d}_{\alpha ,m}=\sum _{\beta =0}^{N-1}\sum _{\alpha =0}^{N-1}{c}_{n,\beta }{b}_{\beta ,\alpha }{d}_{\alpha ,m}.& \left(19\right)\end{array}$
Each element in the a and b matrices have a one-to-one mapping into scan order arrays, and the first element of the scan order array (a′₀=a_0,0) is the following:

$\begin{array}{cc}{a}_o^{\prime }=\sum _{\beta =0}^{N-1}\sum _{\alpha =0}^{N-1}{\mathrm{<figure-callout\; label="c"\; filenames="US07190723-20070313-D00004.png,US07190723-20070313-D00008.png"\; state="\{\{state\}\}">c</figure-callout>}}_{0,\beta }{d}_{\alpha ,0}{b}_{\beta ,\alpha }=\sum _{\gamma =0}^{N^2-1}{\mathrm{<figure-callout\; label="f"\; filenames="US07190723-20070313-D00004.png,US07190723-20070313-D00008.png"\; state="\{\{state\}\}">f</figure-callout>}}_{0,\gamma }{b}_{\gamma }^{\prime }.& \left(20\right)\end{array}$
Each element of f is determined on a term by term basis according to scan order. For example, using the scan order illustrated inFIG. 7A and N=8, b′_o=b_0,0, b′₁=b_0,1, b′₂=b_1,0, . . . b′₆₃=b_7,7, then f_0,0=c_0,0d_0,0, f_0,1=c_0,0d_1,0, f_0,2=c_0,1d_0,0, . . . and f_0,63=C_0,7d_7,0.

In a similar fashion, the elements of the DCT-domain motion compensation (MC) matrix of the (G) are found. In one preferred embodiment, thememory1708 includes a complete set of G matrices to account for all possible integer pixel sub-block placements within the motion compensated block. As those skilled in the art are well aware, MPEG-2 allows for half pixel translations of a sub-block, which are accomplished through a linear combination of integer pixel translations. For the sake of clarity, motion vectors that translate a block of pixels from a reference frame into an inter-coded frame are considered to be integer translations, but those skilled in the art understand half-integer translations, and such translations are considered to be within the scope of the invention.

Motion Compensation

FIGS. 18–20 illustrate exemplary logic that is implemented by thetranscoder134 for applying motion compensation in the run-level domain to frames that are transcoded. InFIG. 18,steps1800 illustrate one embodiment, among others, for applying a motion compensation scheme within thetranscoder134, responsive to thetranscoder134 selectively reducing the bit size of a frame using either requantization or thresholding or both requantization and thresholding. InFIG. 19, non-limitingexemplary steps1900 illustrate one embodiment of the motion compensation scheme. InFIG. 20, non-limitingexemplary steps2000 illustrate one embodiment of accumulating drift, which is introduced by requantization or thresholding or both requantization and thresholding and which is used in the motion compensation scheme illustrated inFIG. 19. InFIGS. 18–20, levels that have been processed by requantization, wherein the quantization parameter has been changed from Q₁to Q₂, and levels that have been processed by thresholding are denoted with primes, e.g., l′, whereas, levels that have not been requantized (change of quantization parameter) or have not been thresholded are not primed, e.g., l.

Refer toFIG. 18, instep1802, theprocessor1702 receives a current frame from theVLD902. The current frame includes, among other things, headers, and sets of quantized run-level pairs denoted by {r, l(Q₁)}. If the current frame is inter-coded, it also includes among other things motion vectors. A consequence of requantization and/or thresholding is a drift in the levels of inter-coded frames. In a conventional transcoder that converts a frame back into pixel domain values, the drift is compensated by performing standard motion compensation on the difference of the pre-transcoded and transcoded reference frames, and performing a DCT of the results. The accumulated drift for a frame is made up of matrices, or equivalently 64 element arrays, of accumulated sub-macroblock drift, and the drift is in the DCT-domain. When an I-picture, the first picture in a GOP, is received, the accumulated drift (D) is set to zero. After the I-picture has been processed by requantization and/or thresholding, the drift for each sub-macroblock, i.e., the difference between the incoming levels and the processed levels, is determined. The accumulated drift (D) is buffered in drift buffers1710 so that it can be used to correct inter-coded frames.

Instep1804, therate controller1704 initializes the parameters used for processing a slice of the current frame. Among other things, therate controller1704 determines the amount of bits to shave off of the current slice and initializes quantization parameters and thresholding parameters. Instep1806, therate controller1704 determines whether the current frame is an I-picture. If the current frame is an I-picture, therate controller1704 proceeds to step1808 and applies motion compensation to the current slice of the current frame. Typically, P-pictures and B-pictures are also requantized as part of motion compensation and will be discussed hereinbelow. After determining that the current frame is an I-picture, or after determining the current frame is not an I-picture and applying motion compensation on the current slice of the current frame, therate controller1704 proceeds to step1810 and determines whether to requantize the slice and whether the current frame is an I-picture. Therate controller1704 proceeds to step1812 only if both conditions are met, i.e., that the current frame is an I-picture and that it should be requantized. As previously described hereinabove, the determination to requantize or not is preferably based upon multiple parameters such as, but not limited, the reduction threshold (R_T), picture type, the maximum initial quantization parameter, and other parameters. As will be explained hereinbelow, if the current frame is a B-picture or P-picture, then as part of the motion compensation performed instep1808 therate controller1704 determines whether to requantize the current slice, and if so, requantizes the current slice. Instep1812, therequantizer1714 requantized the levels using the new quantization parameter Q2.

Instep1814, therate controller1704 determines whether to threshold the current slice. Typically, as previously described, the decision to threshold or not is based in part upon parameters such as the reduction threshold (R_T), the number of bits saved by requantization, and the average of the absolute values of the levels. However, other parameters including fewer parameters, different parameters or more parameters can also be used in the determination for thresholding.

If therate controller1704 decides to threshold the current slice therate controller1704 proceeds to step1816 and thresholds the current slice. In one preferred embodiment, the thresholding is performed using the thresholding logic illustrated inFIG. 15 along with the state machine illustrated inFIG. 16. It should be noted that the levels after thresholding are denoted as L′(Q) (where Q is either Q₁, the initial quantization parameter, or Q₂, the final quantization parameter). If the levels of the current slice were not requantized, then they are functions of Q₁, and if they were requantized, then they are function of Q₂.

After thresholding, or not thresholding, therate controller1704 proceeds to step1818 and determines whether the current frame is a B-picture, and if so proceeds to step1820 and accumulates the drift in the levels caused by requantization and/or thresholding. The drift is accumulated throughout a group of pictures and reset to zero at the beginning of a new group of pictures.

Instep1822, therate controller1704 determines whether the current slice was the last slice of the current frame, and if so, proceeds to step1824. On the other hand, if the current slice is not the last slice of the current frame, therate controller1704 returns to step1804 and continues to process the slices of the current frame until finished.

Instep1824, thescanner938 generates new sets of run-level pairs, which are denoted by {r′, l′(Q)}, and theprocessor1702 updates the accumulated drift if the current frame is a reference frame, e.g., an I-Picture or a P-Picture. The updating of the accumulated drift is done by buffering the current accumulated drift (T), which was calculated instep1820, into the accumulated drift (D). In one preferred embodiment, the requantization and thresholding are done in parallel.

Instep1826, theprocessor1702 sends the processed run-level pairs {r′, l′(Q)} of the current frame to theVLE906 for processing. TheVLE906 converts the run-level pairs into compressed data using Huffman coding and transmits the compressed frame.

Refer toFIG. 19,steps1900 illustrate an exemplary method of applying motion compensation in the DCT-domain. Instep1902, therate controller1704 inverse quantizes the levels of the slice to produce unquantized levels, which are denoted as l.

Instep1904, therate controller1704 extracts DCT domain MC matrices (G) and selected matrices of accumulated drift (D) from thememory1708. Each of the DCT domain MC matrices in thememory1708 is associated with a motion vector, and therate controller1704 uses the motion vectors for the macroblocks of the current slice to determine which DCT domain MC matrices (G) to extract. The selected matrices of accumulated drift correspond to the reference frame sub-macroblocks for motion compensated sub-macroblocks of the current frame, and therate controller1704 uses the header information of the current frame to determine which matrix of accumulated drift (D_i) to extract, where “i” denotes that the matrix of accumulated drift corresponds to the “i^th” block of the reference frame. In other words, motion vectors of the current frame map matrices of accumulated drift to sets of run-level pairs of the current frame. When the current frame is a P-Picture, the matrices of accumulated drift are from the preceding reference frame such as an I-Picture or P-Picture. When the current frame is a B-Picture, the matrices of accumulated drift are selected from the preceding reference frame such as an I-Picture or P-Picture, and from the subsequent P-Picture reference frame.

Instep1906, for each set of run-level pairs {r, l} of the current slice therate controller1704 calculates motion compensation for the accumulated drift by matrix multiplication of the G matrix with the associated matrix of accumulated drift D and the product (GD) is added to the unquantized levels. The product (GD) and the unquantized level (l) are then buffered. Typically, as illustrated inFIG. 19, a block in the current frame receives information from four blocks in a reference frame, and consequently, for a set of levels

$I=I+\sum _{i=1}^4{G}_i{D}_i.$
Conceptually, the G matrix maps accumulated drift associated with a sub-macroblock in reference frame into a motion compensated sub-macroblock of the current frame.

Instep1908, therate controller1704 determines whether the current slice should be requantized. Typically, the decision to requantize or not is performed using the logic previously described hereinabove.

Instep1910, thequantizer1806 requantizes the levels of the current slice using the quantization parameter Q₂. The requantized sets of run-level pairs are denoted by {r, l′(Q₂)}. On the other hand, if therate controller1704 had determined not to requantize the current slice, then instep1912, thequantizer1806 requantizes the levels of the current slice using the quantization parameter Q₁. After the unquantized levels have been converted back into quantized levels, l(Q), therate controller1704 is done with motion compensation.

FIG. 20 illustrates exemplary steps taken by therate controller1704 to accumulate drift. Instep2002, therate controller1704 inverse quantizes the processed levels (l′(Q)) to produce unquantized processed levels (l′) for each set of levels in the current slice. If the current frame is an I-Picture, then therate controller1704 inverse quantizes the initial quantized levels (l′(Q₁)) to produce unquantized non-processed levels (l) for each set of levels in the current slice. However, if the current frame is not an I-Picture, then unquantized non-processed levels (l) were produced and buffered when motion compensation was applied instep1906. In that case, the unquantized non-processed level (l) of the current slice are extracted from thememory1708.

Instep2004, therate controller1704 calculates the current accumulated drift that is associated with each set of run-level pairs in the current slice, and buffers the current accumulated drift in a temporary array (T). The current accumulated drift is sum of the motion compensation of the accumulated drift, (GD), from prior reference frames, plus the instantaneous drift, the difference in the unquantized non-processed levels (l) and the unquantized processed levels (l′), i.e.,

$\mathrm{Drift}=\left(1-1^{\prime }\right)+\sum _{i=1}^4{G}_i{D}_i.$
The accumulated drift from prior reference frames (D) is not updated until the entire current frame has been processed, so that the accumulated drift does not include artifacts of the current frame.

In one preferred embodiment, thememory1708 includes buffers for at least two frames worth of drift so that it can include drift for both the immediately preceding reference frame (an I-Picture or P-Picture) and the current reference frame (a P-Picture) in order to properly process B-Pictures.

In one preferred embodiment, the drift for different types of frames such as video-frames, top video-fields, and bottom video-fields are accumulated inmemory1708 separately. In this embodiment, theprocessor1702 determines whether the current frame is a video-frame, i.e., non-interlaced, or a top video-field, or a bottom video-field using the header information of the current frame and then extracts the appropriate sets of drift from thememory1708 for motion compensation and updates the appropriate sets of drift. It should be emphasized, that for the sake of clarity, the steps of motion compensation were described in a sequential manner. However, as those skilled in the art will recognize, the steps could be implemented in a different order and/or in parallel. In one preferred embodiment, steps such as, but not limited to, quantizing, inverse quantizing, calculation of new run values, and linear operations of matrices are done in parallel in enhance computational efficiency.

In one preferred embodiment, B-Pictures are processed without motion compensation. In other words, for a B-Picture steps1900 are skipped over. Motion compensation of B-Pictures can be skipped because B-Pictures are not used as reference pictures and any drift error in the B-Pictures is not accumulated and, consequently, is used in the motion compensation of subsequent pictures. Since many MPEG-2 streams contain a majority of B-Pictures, computational efficiency is enhanced by not doing motion compensation for B-Pictures.

Although exemplary preferred embodiments of the present invention have been shown and described, it will be apparent to those of ordinary skill in the art that a number of changes, modifications, or alterations to the invention as described may be made, none of which depart from the spirit of the present invention. Changes, modifications, and alterations should therefore be seen as within the scope of the present invention. It should also be emphasized that the above-described embodiments of the present invention, particularly, any “preferred embodiments” are merely possible non-limiting examples of implementations, merely setting forth a clear understanding of the principles of the inventions.

Claims (49)

1. A method of transcoding a digital stream of compressed frames, the method comprising the steps of:

(a) receiving a compressed video frame having content information and non-content information included therein;

(b) determining the total compressed size of the frame (N_T);

(c) determining from at least the total compressed size of the frame the total number of bits (N_S) to shave from compressed frame;

(d) determining a plurality of statistics about a given portion of the frame;

(e) determining whether to requantize the given portion based at least in part upon at least one of the statistics about the given portion;

(f) responsive to determining to requantize the given portion, requantizing the levels of the given portion;

(g) determining whether to threshold the given portion based at least in part on the at least one of the statistics about the given portion;

(h) responsive to determining to threshold the given portion, thresholding the levels of the given portion;

(i) transmitting the given portion; and

prior to step (h),

determining a target number (N_THRESH) of levels to set to zero by thresholding;

determining the number of levels (CNT) beneath a threshold function having a predetermined width and height profile;

responsive to the number of levels (CNT) beneath the threshold function being within a predetermined range of the target number (N_THRESH), setting the magnitude of the levels beneath the threshold function to zero; and

responsive to the number of levels (CNT) beneath the threshold function being outside of the predetermined range of the target number (N_THRESH), adjusting the threshold function.

2. The method ofclaim 1, prior to step (d), further including the steps of:

(j) decompressing the video frame into a DCT-domain, wherein in the DCT-domain blocks of pixel information are represented as blocks of levels;

(k) parsing the frame into a plurality of portions; and

prior to step (i), re-compressing the given portion, wherein the given transmitted in step (i) is the re-compressed given portion.

3. The method ofclaim 1, wherein the plurality of statistics about the given portion include the compressed content size (Ssize) of the given portion.

4. The method ofclaim 1, wherein the plurality of statistics about the given portion include the average of the quantization parameters used to quantize the levels of the given portion.

5. The method ofclaim 1, wherein the plurality of statistics about the given portion include the average of the runs in the given portion.

6. The method ofclaim 1, further including the steps of:

prior to the step (e), determining the initial compressed content size (Sint) of the given portion;

prior to the step (e), determining a target number of bits to shave (N_shave) from the given portion;

prior to the step (i), determining the final compressed content size (Sfnl) of the given portion; and

calculating the reduction error (e) for the given portion, wherein the reduction error is defined as the difference between the target number of bits to shave and the difference between final and initial compressed content size, e=N_shave−(Sfnl−Sint).

7. The method ofclaim 6, further including the step of:

accumulating the reduction error, wherein the accumulated reduction error is used in determining the number of bits to shave from a subsequent portion of the frame.

8. The method ofclaim 6, wherein the number of portions in the frame is N_portions, wherein the given portion is the K^thout of the N_portionsto be processed according to steps (e), (f), (g) and (h), and further including the steps of:

determining the total compressed content size (C_T) of the frame; and

accumulating the reduction error for the preceding (K−1) portions;

wherein the number of bits to shave from the given portion is given by

$N_{\mathrm{SHAVE}}=S_{\mathrm{INT}}\times \frac{N_S}{C_T}+\frac{E}{N_{\mathrm{portion}}-K+1},$

where E is the accumulated reduction error for the preceding portions.

9. The method ofclaim 1, further including the step of:

after step (f) and prior to step (g), re-determining at least one of the statistics of the given portion, wherein the at least one re-determined statistics about the given portion is used in step (g).

10. The method ofclaim 9, wherein the determined statistics include the compressed content size (Ssize) of the given portion, and the compressed content size is redetermined after the given portion has been requantized.

11. The method ofclaim 1, further including the steps of:

determining a requantization parameter for the given portion based at least upon one of the determined statistics about the given portion.

12. The method ofclaim 11, wherein the given portion was previously quantized using a first quantization parameter (Q₁), wherein the requantization parameter (Q₂) is twice the first quantization parameter, Q₂=2Q₁.

13. The method ofclaim 11, wherein the given portion was previously quantized using a first quantization parameter (Q₁), wherein the requantization parameter (Q₂) is four times the first quantization parameter, Q₂=4Q₁.

14. The method ofclaim 11, wherein the requantization parameter is chosen such that the compressed content size of the given portion is reduced by approximately 60% to 70%.

15. The method ofclaim 11, wherein the requantization parameter is chosen such that the compressed content size of the given portion is reduced by at least 70%.

16. The method ofclaim 11, wherein the given portion is made up of multiple blocks of previously quantized levels, and step (d) includes determining both the maximum quantization parameter (Q1MAX) used in quantizing the previously quantized levels and the average of the previously quantized levels (Lavg), and wherein Q1MAX and Lavg are used in determining the requantization parameter.

17. The method ofclaim 11, further including the step of:

determining the picture-type of the video frame, wherein the video frame is an MPEG frame, and the group of picture-types consist of I-Picture, P-Picture and B-Picture, and wherein the picture-type of the video frame is used in determining the requantization parameter.

18. The method ofclaim 1, wherein the step of adjusting further includes the step of:

changing the height profile of the threshold function.

19. The method ofclaim 18, wherein the height profile is increment.

20. The method ofclaim 18, wherein the height profile is decrement.

21. The method ofclaim 1, wherein the levels are arranged in order of scan position, arid the threshold function extends over a first range of scan positions, and the step of adjusting further includes the step of:

shifting the relative location of the threshold function such that the threshold function extends over a second of range of scan positions.

22. The method ofclaim 21, wherein the smallest scan position included in the first range of scan positions is smaller than the smallest scan position included in the second range of scan positions.

23. The method ofclaim 21, wherein the smallest scan position included in the first range of scan positions is greater than the smallest scan position included in the second range of scan positions.

24. The method ofclaim 1, wherein the threshold function extends over a first range of scan positions, and the given portion is made up of multiple blocks of levels, and the step of adjusting further includes the steps of:

(l) partitioning the blocks of the given portion into a first group and a second group of blocks;

associating a second threshold function with the second group, wherein the second threshold function extends over a second range of scan positions;

(m) determining the number of levels (CNT1) beneath the first threshold function;

(n) determining the number of levels (CNT2) beneath the second threshold function;

(o) responsive to the sum of CNT1 and CNT2 being within the predetermined range of the target number (N_THRESH), setting the magnitude of the levels beneath the first and second threshold functions to zero;

(p) responsive to the sum of CNT1 and CNT2 being outside of the predetermined range of the target number (N_THRESH), transferring a predetermined number of blocks from the first group to the second group; and

responsive to the sum of CNT1 and CNT2 being outside of the predetermined range of the target number (N_THRESH), repeating steps (m), (n), (o), and (p) until the sum of CNT1 and CNT2 is within the predetermined range of the target number (N_THRESH).

25. The method ofclaim 1, wherein the given portion of the frame is a slice of the frame having multiple blocks of levels, and the blocks of levels are thresholded in parallel.

26. An apparatus in a network for transcoding a digital stream of compressed frames, the apparatus comprising:

a decoder adapted to decompress a video frame having content information and non-content information included therein into the DCT-domain, wherein the content-information carried by the frame is represented as levels in the DCT-domain;

a rate controller adapted to receive the video frame, parse the video frame into a plurality of portions, determine a plurality of statistics about a given portion, and determine a target number of bits to shave (N_shave) from the given portion;

a requantizer adapted to requantize levels of the given portion;

a thresholder adapted to threshold the given portion, wherein the rate controller determines whether the given portion should be requantized and whether the given portion should be thresholded based at least in part on at least one of the statistics about the given portion; and

an encoder adapted to compress the frame, wherein the compressed size of the frame is approximately the same as a target size;

wherein the rate controller is further adapted to determine a target number (N_THRESH) of levels to set to zero by thresholding; wherein the thresholder is adapted to determine the number of levels (CNT) beneath a threshold function having a predetermined width and height profile, responsive to the number of levels (CNT) beneath the threshold function being within a predetermined range of the target number (N_THRESH), the thresholder sets the magnitude of the levels beneath the threshold function to zero, and responsive to the number of levels (CNT) beneath the threshold function being outside of the predetermined range of the target number (N_THRESH), the thresholder adjusts the threshold function.

27. The apparatus ofclaim 26, wherein the plurality of statistics about the given portion include the compressed content size (Ssize) of the given portion.

28. The apparatus ofclaim 26, wherein the plurality of statistics about the given portion include the average of the quantization parameters used to quantize the levels of the given portion.

29. The apparatus ofclaim 26, wherein the plurality of statistics about the given portion include the maximum quantization parameter used to quantize the levels of the given portion.

30. The apparatus ofclaim 26, wherein the rate controller determines the initial compressed content size (Sint) of the given portion and the final compressed content size (Sfnl) of the given portion, and the rate controller calculates the reduction error (e) for the given portion, wherein the reduction error is defined as the difference between the target number of bits to shave and the difference between final and initial compressed content size, e=N_shave—(Sfnl—Sint).

31. The apparatus ofclaim 30, wherein the rate controller accumulates the reduction error, and the accumulated reduction error is used in determining the number of bits to shave from a subsequent portion of the frame.

32. The apparatus ofclaim 30, wherein the number of portions in the frame is N_portions, wherein rate controller processes the plurality of portions sequentially and the given portion is the K_thportion out of the N_portionsto be processed, the rate controller further adapted to determine the total compressed content size (C_T) of the frame, and accumulate the reduction error for the preceding (K−1) portions, wherein the number of bits to shave from the given portion is given by

$N_{\mathrm{SHAVE}}=S_{\mathrm{INT}}\times \frac{N_S}{C_T}+\frac{E}{N_{\mathrm{portion}}-K+1},$

where E is the accumulated reduction error for the preceding portions.

33. The apparatus ofclaim 26, wherein the rate controller re-determines at least one of the statistics of the given portion responsive to requantizing the given portion, wherein the at least one re-determined statistics about the given portion is used to determine whether to threshold the given portion.

34. The apparatus ofclaim 33, wherein the determined statistics include the compressed content size (Ssize) of the given portion, and the compressed content size is re-determined after the given portion has been requantized.

35. The apparatus ofclaim 26, wherein the rate controller is further adapted to determine a requantization parameter for The given portion based at least upon one of the determined statistics about the given portion.

36. The apparatus ofclaim 35, wherein the given portion was previously quantized using a first quantization parameter (Q₁), wherein the requantization parameter (Q₂) is twice the first quantization parameter, Q₂=2Q₁.

37. The apparatus ofclaim 35, wherein the given portion was previously quantized using a first quantization parameter (Q₁), wherein the requantization parameter (Q₂) is four times the first quantization parameter, Q₂=4Q₁.

38. The apparatus ofclaim 35, wherein the requantization parameter is chosen such that the compressed content size of the given portion is reduced by approximately 60% to 70%.

39. The apparatus ofclaim 35, wherein the requantization parameter is chosen such that the compressed content size of the given portion is reduced by at least 70%.

40. The apparatus ofclaim 35, wherein the given portion is made up of multiple blocks of previously quantized levels, and the rate controller determines both the maximum quantization parameter (Q1MAX) used in quantizing the previously quantized levels and the average of the previously quantized levels (Lavg), and wherein Q1MAX and Lavg are used in determining the requantization parameter.

41. The apparatus ofclaim 35, wherein the rate controller is further adapted to determine the picture-type of the video frame, wherein the video frame is an MPEG frame, and the group of picture-types consist of I-Picture, P-Picture and B-Picture, and wherein the picture-type of the video frame is used in determining the requantization parameter.

42. The apparatus ofclaim 26, wherein the thresholder adjusts the threshold function by changing the height profile of the threshold function.

43. The apparatus ofclaim 42, wherein the height profile is increment.

44. The apparatus ofclaim 42, wherein the height profile is decrement.

45. The apparatus ofclaim 26, wherein the levels are arranged in order of scan position, and the threshold function extends over a first range of scan positions, and the thresholder adjusts the thresholding function by shifting the relative location of the threshold function such tat the threshold function extends over a second of range of scan positions.

46. The apparatus ofclaim 45, wherein the smallest scan position included in the first range of scan positions is smaller than the smallest scan position included in the second range of scan positions.

47. The apparatus ofclaim 45, wherein the smallest scan position included in the first range of scan positions is greater than the smallest scan position included in the second range of scan positions.

48. The apparatus ofclaim 26, wherein the threshold function extends over a first range of scan positions, and the given portion is made up of multiple blocks of levels, and the thresholder adjusts the thresholding fraction by the steps of:

partitioning the blocks of the given portion into a first group and a second group of blocks;

associating a second threshold function with the second group, wherein the second threshold function extends aver a second range of scan positions;

(a) determining the number of levels (CNT1) beneath the first threshold function;

(b) determining the number of levels (CNT2) beneath the second threshold function;

(c) responsive to the sum of CNT1 and CNT2 being within the predetermined range of the target number (N_THRESH), setting the magnitude of the levels beneath the first and second threshold functions to zero;

(d) responsive to the sum of CNT1 and CNT2 being outside of the predetermined range of the target number (N_THRESH), transferring a predetermined number of blocks from the first group to the second group; and

responsive to the sum of CNT1 and CNT2 being outside of the predetermined range of the target number (N_THRESH), repeating steps (a), (b), (c), and (d) until the sum of CNT1 and CNT2 is within the predetermined range of the target number (N_THRESH).

49. The apparatus ofclaim 26, wherein the given portion of the frame is a slice of the frame having multiple blocks of levels, and the blocks of levels are thresholded by the thresholder in parallel.

Images